Comment by averne_

No problem, just be aware there's a bunch of optimizations I haven't had time to implement yet. In particular, I'd to remove the reset kernel, fuse the VLD/IDCT ones, and try different strategies and hw-dependent specializations for the IDCT routine (AAN algorithm, packed FP16, cooperative matrices).